Biomedical & Pharmacology Journal, June 2022. Vol. 15(2), p.

699-716

Brain Tumor Segmentation and Classification in MRI

using Clustering and Kernel-Based SVM
Anil Kumar Mandle*, Satya Prakash Sahu and Govind Gupta

Department of Information Technology, National Institute of Technology, Raipur, India.

Corresponding Author E-mail: [email protected]

https://dx.doi.org/10.13005/bpj/2409

(Received: 28 June 2021; accepted: 3 June 2022)

Brain tumor and other nervous systems cancer are one of the leading causes of death
for many patients. Magnetic resonance imaging (MRI) is the most important medical imaging
modality for diagnosing brain tumors and other disorders in the brain. Manual evaluation
of several MRI images by radiologists or experts for diagnosing brain tumors especially at
early stages is a challenging task. Hence, this paper proposes an automated framework for
the segmentation and classification of brain tumors using K-means clustering and kernel-
based support vector machine (K-SVM). The major steps of the proposed framework consist
of preprocessing, segmentation, feature extraction with selection, and classification. In the
preprocessing step, the regions of interest (ROI) are extracted using skull stripping and a median
filter. In the next step, the tumor is segmented using an optimized K-means algorithm. Further,
discrete wavelet transform (DWT)-based texture features are used for feature extraction, and
significant features are selected by principal component analysis (PCA). Finally, the kernel-
based support vector machine (K-SVM) is used for the classification of brain tumor types into
benign and malignant, with a dataset using 160 MRI images, consisting of 20 normal and 140
abnormal. Experimental findings demonstrated the efficacy of the proposed framework with
98.75% accuracy, 95.43% precision, and 97.65% recall. The simulation findings emphasize
the importance of the proposed system as compared to state-of-the-art techniques in terms of
coherence parameters and performance.

Keywords: Brain tumor MRI; DWT; K-means algorithm; KSVM; PCA; Segmentation.

Brain tumors might be a dangerous disease various parts of the brain rather than Ultrasound or
for a human. It is a mass or growth of abnormal CT scan and X-ray1. In recent years, the medical
cells in the brain. In different imaging modalities sector has benefited from the implementation of IT
such as Single-photon Emission Computed and e-health care systems, which allow healthcare
Tomography (SPECT), X-ray, Positron Emission experts to deliver quality health care to patients.
Tomography (PET), Computed Tomography (CT), Kumari et al2 studied brain MRI classification
and Magnetic Resonance Imaging (MRI), the through feature extraction technique followed by
MRI has become a very useful medical diagnostic the Support Vector Machine (SVM) classifier to
tool for the identification and diagnosis of brain categorize the tumor.
tumors. In addition, MRI provides clear images of

This is an Open Access article licensed under a Creative Commons license: Attribution 4.0 International (CC-BY).
Published by Oriental Scientific Publishing Company © 2022
700 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

The most popular grading system from algorithm overcomes the difficulties present in the
Grade-I to Grade-IV of brain tumors distinguishes traditional approach.
between benign and malignant tumor types. Benign • Discrete Wavelet Transform (DWT)-based
tumors are classified as grade I and II glioma, texture features are used to feature extraction and
while malignant tumors are classified as grade significant features are selected by the principal
III and IV glioma. Medical imaging modalities component analysis (PCA). This approach is
are segmented to diagnose compromised tumor mainly used to reduce complexity and increase
tissues3. The method of dividing an image into accuracy.
separate regions or blocks that share similar and • The K-SVM-based algorithm is proposed for
equivalent properties like color, texture, contrast, tumor classification. K-SVM overcomes the
brightness, borders, and gray level is known as difficulties present in the traditional SVM in which
segmentation. Using MR images or other imaging accuracy is further improved.
modalities, from tumor cells like gray matter (GM), The remaining section of the paper
brain tumor segmentation identifies tumor tissues is structured as follows: Section 2 addresses
such as dead and edema cells GM, white matter related works, Section 3 described materials and
(WM), and cerebrospinal fluid (CSF), as well as techniques, including proposed method, Section 4
normal brain tissues. It is apparent that if a tumor- Experimental results and discussion, and Section 5
infected patient’s chances of survival are greatly involve conclusions and future work scope.
improved when the tumor is diagnosed correctly Related works
in the early stages4. Therefore, the solution for the Segmentation and classification of brain
best result, the discussing of brain tumors using tumors have been carried out in various studies and
types of images has been gained importance to the the most common methods used in these articles
department of radiology. are based on classification, clustering algorithms,
Although there exist numerous methods thresholding, and morphological operations.
for brain tumor classification, they have certain In5, for segmentation, Fuzzy c-means has been
limitations which require being precise while utilized and for feature extraction, a gray level
effective with brain tumor segmentation and length matrix is used, and MRI brain images are
classification. The first problem in these methods is labeled using the SVM technique with separate
their binary tumor categorization, which leaves the kernels. The accuracy results in performances of
radiologist with various questions. The reason for SVM classification for kernel functions including
this is that a radiologist’s categorization of a tumor Linear, Quadratic, and Polynomial. K-means and
as benign or malignant is difficult to determine the Histogram Normalization are segmentation
treatment and prevention options for the patient. operations. De-noising methods such as the
For radiologist clarity and comprehension, the median filter, wiener filter, averaging filter, spatial
classification must be multi-class, which classifies filter, adaptive filter, and Gaussian filter are used
brain tumors into relevant categories. Furthermore, to eliminate noise from the MRI input image
researchers have a significant problem in achieving during preprocessing. The resultant will then be
exact findings because of a lack of data. To address normalized before being fed into a classifier for
these key limitations, we proposed a kernel-based classification 6. After segmentation, the tumor
support vector machine (K-SVM) is applied for is identified using the k-means algorithm. The
the classification of brain tumor types benign and magnetic resonance image is carried out using the
malignant. NB classifier and SVM classifier to achieve more
The contribution of the proposed accuracy in estimation and classification. Naive
methodology is presented below Bayes was used for the classifier, while the best
• Preprocessing is the initial step for the median accuracy of the SVM classifier. In comparison
filter to be utilized in which noises present in the to other classifiers, the SVM classifier performs
images are removed. well. Since the tumor area is not exact and reliable,
• Segmentation is performed based on the K-means the suggested approach has certain drawbacks
algorithm. The proposed optimized K-means on the boundary.SVM is used for classification.
 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 701

Here, the input images must go through a series degree13. This final value of equivalence degree is
of steps before being focused on the region of used to evaluate the segmentation of images and
interest (ROI). The SVM is used for classification combine regions. The method has the limits up to
in this system7. It is possible to attain a degree of single MRI images of brain and CT, artificially
precision. The best way to segment and identify a generate images. From the above-mentioned
tumor in brain MRI images is to use CNN-based survey, it has been justified that some techniques
image segmentation, pre-processing procedures, are designed for segmentation and classification,
and data augmentation. Bias field improvement, some are simply for feature extraction, and some
intensity, and regular image patches are all part of are only for classification.
it. To research deep structures, this device is based Proposed methodology
on convolution layers 3×3 with kernels8. A strategy This section presents a detailed work of
is planned in9 a method and classified the MRI brain the proposed framework in which an optimized
tumor where the input is 3-D MR images. This K-Means algorithm14 and K-SVM are used for
method created high-quality results for accuracy, efficient segmentation and classification of the
sensitivity, and specificity. For a multiclass brain MRI brain tumors. The major step of this work
tumor, a procedure has based on segmentation consists of preprocessing, segmentation, feature
and feature extraction for images was conferred. extraction with selection, and classification. In
A total of 428 MR images were used in this study. the preprocessing step, the regions of interest
The ANN and PCA-ANN methods are used in (ROI) are extracted using skull stripping and the
classification to increase precision10. The explained noise effects are removed by using the median
how to solve the image visibility and clearance issue filter. In the next step, the tumor is segmented
with the fewest possible configuration parameters using an optimized K-means algorithm. Further,
on the input image11. They did this by dividing DWT-based texture features are used for feature
the medical specialty procedure into two groups extraction, and significant features are selected by
of automated procedures. Inset one, they focus the Principal Component Analysis (PCA). Finally,
on image quality and segmentation of the object the kernel-based support vector machine (K-SVM)
of interest, which is a tumor, by forming an edge is applied for the classification of brain tumor
diagram. In the second package, they performed types into benign and malignant. Fig. 1 illustrates
data analysis by measuring the parameters derived the block diagram of the proposed framework in
from the diagnosis. The multi-modal hybrid fusion which a sequence of operations is performed such
and convolution neural network recognition system as pre-processing of brain MRI images, image
was combined in this experiment. The use of the enhancement to increase contrast and brightness,
2D CNN and 3D CNN multimodal extension by skull stripping operation, segmentation, feature
having brain lesion with various characteristics in extraction, feature selection, and classification
3D was used in12 this article. By solving the 2D using K-SVM.
CNN for various modal details at raw input faults to Pre-processing Steps
boost convergence speed, a real normalization layer In this step, the regions of interest
was added to the convolution and pooling layers (ROI) are extracted using skull stripping and
to remove the issue of overfitting. Resulting in the median filter. It improves the accuracy of raw MR
improvement of the loss function, so weighted loss images. Furthermore, pre-processing aids in the
function is added in the lesion area to develop the improvement of certain MR image parameters,
feature learning. The image segmentation method such as the signal-to-noise ratio, the removal
uses the FCM clustering algorithm through a rough of irrelevant noise and unwanted bits in the
set theory. The authors construct the feature values background, the smoothing of the inner part of
obtained from the FCM segmentation outcome the area, and the preservation of its edges15. To
and the image is separated into small areas based increase the signal-to-noise ratio in our proposed
on the attribute. Biased values are obtained with framework.
value decrease and used for the calculated variation Median filter
between the region and comparison of the region. It is a valuable tool for separating out-of-
Later was realized through similarity variation range isolated noise from valid image features like
702 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

edges and lines to some degree. The median filter stripping is a critical step. It distinguishes between
replaces a pixel with the median, rather than the the skull, fat, muscle, and the part of the brain
average, of all pixels in a neighborhood. that is not of concern. It eliminates the extra parts
The median filter operates with replacing of the brain to make the process of segmentation
the value of a pixel with the middle value of the simpler and more efficient. The skull stripping
pixels in a little window adjacent to it. For the n×n (SS) algorithm 16, which relied on morphological
pixel window, the median filtering operation is operations, created poor skull stripping and was
represented by the following representation: ineffective when applied to DICOM images. The
purpose of skull striping for blob identification
and marking system is used in this article. A blog
is an area in an image where all the pixels have the
same or almost the same properties. All the points
...(1) in a blob are regarded to be similar in some way
and different from their surroundings. The biggest
where i(p,q)-Pixel value at a point (p,q), p,q ∈ Z. benefit of using blob over the edge and corner
The threshold is a significant and essential detectors is that it produces improved performance
approach to pixel-based segmentation. It is used to and support to provide detailed nearest regions in
transform grayscale pictures into binary images an image more precisely.
in their most basic form. For segmentation, a Algorithm: Skull Stripping
threshold value is selected. Pixels with an intensity Begin
value less than or equal to the threshold value are Step1: Initialize Image Iij and Q[n]
changed to black, while those with an intensity Step2: set current label to 1 i.e. curlab=1
value greater than or equal to the threshold value Step3: if (Iij ==foreground && !=labeled)
are converted to white. Using Otsu’s technique, the {
threshold value is determined and converted into a Step4:for(i=2, i<n, i++)
gray image to a binary image. Step5: labeled= curlab // labeling of pixel
Image converted get as follow: after first pixel.
Step6:Iij = Qi-1
}
else (Iij == background)
Step7: for all Q[n]
...(2) Implement breadth-first search or depth-first
where, b0 and b1 are for two classes Search algorithm (to extract neighbor of Q[n])
probabilities and divided by a threshold value tv for (k=0,k=n,k++)
where, a20 and a21 are the variances two if (Qk+1=foreground&& Qk+1 !=labeled)
classes for the calculate the maximum value of Set Qk+1= curlab
a2
0
(tv) else (Qk ==background)
Step8: apply step 3 for all pixels in the Image
Segmentation
Image Segmentation is a procedure of
...(3) segregating an image into a collection of connected
sets of pixels i.e., into regions, linear structures,
and 2Dshapes. Image Segmentation is the precise
part of the clinical diagnosis. Depending on the
...(4) image, segmentation can be a very sophisticated
where, b0 (tv) and b1(tv) calculated values process. The segmentation of the infested regions
by the L histogram. of the MRI brains is attained by the subsequent
stages: In the primary step, the pre-processed MR
Skull Stripping (SS) brain image is a binary image converted by the
For segmentation for brain tumors, skull Otsu Thresholding technique17. Then in the next
703 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

stage, the segmentation process will be applied

to the binary image. In this study, Enhanced K-means clustering algorithm objective function
K-means is used for effective segmentation of the as follows:
brain. The K-means machine learning technique
is the best among the most popular partitioning
method. In this method, the data is separated into
a collectively k finite numeral of clusters and for
the allocated each sample it gives the clustered ...(5)
index. In the case of massive data, the K-means
clustering is highly deserved than the Hierarchical where,
clustering. The clustering technique K-means is K is the number of clusters, n is the number of
commonly applied for the unsupervised clustering cases, xi is the Case(i), cj –Centroid cluster j, and
procedure. Understanding the algorithm is easy and || xi(j) - cj ||2 and is the distance function.
a stronghold when making it practice. But it does
not provide valid accuracy. So, in this paper, we Algorithm: K-Means Algorithm
proposed Enhanced K-means i.e., selecting initial Step 1: MSE= large number
random centers with précising probabilities which Step 2: select initial cluster centroids {cj}jk=1;
increases the speed and also the correctness of the Step 3: Do
K-means. OldMSE=MSE;

Fig. 1. Block diagram of proposed methodology

 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 704

MSE1=0; Distance d2 (xi, cj);

for j=1 to k endfor
cj==0; nj=0; Step 5: calculate the closest centroid xi to cj
endfor cj = cj +xi; nj=nj+1;
for i=1 to n Step 6:MSE1=MSE1+d2 (xi, cj);
For j=1 to n endfor
Step 4:Calculate squared Euclidean for j=1 to k

Fig. 2. (a) Original image (b) Filtering image.

Fig. 3. (a) After filtering (b) after dilation (c) segmented image

Fig. 4. DWT decomposition 2-levels of an image

705 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

Step 7:nj=max(nj, 1); cj=cj/nj; top of the image to extract the tumor region. The
endfor dilation works by taking two types of input data.
Step8:MSE=MSE1; First, it takes the image to be dilated, and secondly
while (MSE<OldMSE) structuring element, also known as kernel 21.
Pre-processing of the image, extraction of The dilation operation adds pixels to the object
the tumor or ROI, classification of the tumor or that boundary; the pixels adjoined from the object
area, and finally, post-processing are all steps in the depend upon the shape and size of the kernel22.
segmentation of MRI images18. K-means algorithm The final output pixel value will be the maximum
using the improving for the segmentation of low pixel value of all the input pixels. We have used
contrast medical images19, can save time. KSVM dilation since it grows in size and thickens while
can be used to perform region-based classification. the holes in the region will become smaller. Once
Morphological Operations this process is successful the tumor region in the
It is used to separate the tumor and the image will be visible (white), the background will
normal region. Morphological operations depend be black. Masking works by assigning some of
on the pixel’s value, not on the numerical value. the pixels in the image to zero (black background)
It is used to detect the structure of the element20. while the rest (kernel) is non-zero23. Once the
The structure and size of the element will depend morphology dilation has been successful, we
on the value zero or one that is represented in the have taken two input data (images) for masking.
matrix form. The binary or high contract pictures First, we have taken the image from the output
mostly have two types of shading that are white and image of the morphological operation. Secondly,
black. Commonly white represents the foreground the preprocessed MRI image is considered and on
image, and black represents the background of the successful subtraction from the first and second
image. Numerically it is denoted as ‘0’ represents image, a masked image is obtained. Thus, the
black and ‘1 or 255’ used to represent the white masked image further helps in classification
region. Fig. 2 and Fig. 3 show Original and filtering and gives accurate accuracy. Fig. 3 shows the
images, respectively. Morphology dilation operation.
Morphology dilation Feature Extraction
Once the de-noised image has been Feature extraction is the method of
obtained, the morphology dilation is applied on gathering an image’s higher-level details, like
structure, color, contrast, and texture. Texture
analysis is a significant component of both human
visual analysis and machine learning techniques.
By choosing notable features, it is successfully used
to enhance the precision of the diagnostic system.
Discrete Wavelet Transform (DWT)
It is a linear transformation that converts
one data vector into another. The transformed
vector’s length is an integer to the second
power.DWT divides data into four frequency
components24. The four levels are Low Low (LL),
Low High (LH), High Low (HL), and High High
(HH). Many frequency parameters were tested with
the appropriate resolution for their respective sizes.
Digital filtering algorithms are used to represent
the digital signal in time-scale in DWT. The
examined signal is allowed to move through filters
with varying cutoff frequencies and scales. This
signal is compared to the contained basis utilities,
which are described as removed and scaled types
of any previously set of wavelets. Wavelets get
Fig. 5. The implementation of PCA
706 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

Fig. 6. (a) Input original image, (b) Threshold for inputted binary conversion,
(c) morphological operations, and (d) Segmented image are infected area shows dark white.
 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 707

the advantage of providing localized frequency transform is particularly useful because it provides
information on a function, such as: knowledge nearby the signal in both the frequency
and time domains. Transforming images from the
3-Dfield into the occurrence field. Using DWT,
we may fragment an image into levels through
their appropriate DWT coefficients. The DWT is
...(6) performed using flowed decomposition method,
The coefficients of di,j refer to the nearest with low pass and high pass filters that must fulfill
components, while di,j refers to the element specific criteria. Figure 4 depicts the simple DWT
attribute in signal p(s) referring to the wavelet fetidness scheme and its use in MR images. The
function. There are two High-pass and low- coefficients of the high-pass as well as low-pass
pass filter coefficients (s) and g(s), respectively, filters are represented by the functions h(n) and
while wavelet scale and translation variables are g(n), respectively. Therefore, at each scale grade
represented by the parameters i and j. level, there are four levels of imagining (LL,
The discrete wavelet transform (DWT) is LH, HH, HL).DWT for the decomposition at the
used in this analysis to remove features from MRI next scale, individual the sub-band LL is used for
images. The wavelet25 is an efficient mathematical function extraction.
method for extracting functions. The wavelet

Ttable 1. Feature for randomly selected benign 10 images

Images means Standard Entropy Energy Homogen Contrast Correlation

deviation

Img1 0.00282896 0.0897701 3.62834 0.737835 0.927359 0.215517 0.0950755

Img2 0.0031107 0.0897608 3.17346 0.7621 0.935159 0.208843 0.199005
Img3 0.00348476 0.0897471 3.52392 0.740242 0.926743 0.251669 0.073405
Img4 0.00244646 0.0897814 3.45522 0.747275 0.930409 0.221635 0.134566
Img5 0.0042363 0.0897147 3.03266 0.751967 0.929774 0.264739 0.134791
Img6 0.0020681 0.0897909 3.51816 0.769087 0.936531 0.224972 0.0991065
Img7 0.00341361 0.0897498 3.67835 0.752234 0.930798 0.227197 0.0908294
Img8 0.00348021 0.0897473 3.66808 0.744601 0.929468 0.223026 0.0964256
Img9 0.00234208 0.0897842 3.62596 0.753045 0.932837 0.210512 0.0969614
Img10 0.00229851 0.0897853 3.23654 0.748017 0.932026 0.217186 0.144154

Fig. 7. The comparison using parameters mentioned in Table 1

708 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

Principal Component Analysis (PCA) Mean (M)

PCA is a mathematical procedure A calculated picture is created by
for obtaining a reduced number of ages in the multiplying all images’ pixel values with the total
knowledge lattice and is useful for obtaining the number of pixels into the image, and it is expressed
highest possible values of variance. The entire as
approach is based on the subject, which generates
a large dataset; the database is then analyzed for
relations between the set’s independent points.
Its basis vectors are the covariance matrix ...(7)
Eigenvectors of the input data. This can be useful Where n and m are the dimensions of the
when investigating multivariable data processing image a lower value means that the image seems
as new dimensions, referred to as principal to have a good amount of noise removed.
components PCs in the dataset. By choosing Standard Deviation (SD)
the PCs with the highest Eigenvalues, a reduced Calculated for the second step central
dimension can be obtained 26 fig.5 shows the moment there are defines an observed population’s
implementation of PCA. probability distribution and can be used to calculate

Table 2. Feature for randomly selected malignant 10 images.

Images Means Standard Entropy Energy Homogen Contrast Correlation

deviation

Img1 0.00346216 0.089748 3.497 0.750879 0.928926 0.244994 0.0948681

Img2 0.00353077 0.0897453 2.75135 0.776902 0.936323 0.299499 0.0910863
Img3 0.00440501 0.0897066 3.18563 0.768842 0.93502 0.25 0.147433
Img4 0.00458293 0.0896977 3.54839 0.731029 0.924625 0.243882 0.107227
Img5 0.00438539 0.0897217 3.37818 0.763092 0.933862 0.265851 0.143025
Img6 0.00440507 0.0897069 3.18565 0.768845 0.93504 0.2576 0.147435
Img7 0.00476278 0.0899883 3.67655 0.755653 0.931929 0.239989 0.0666697
Img8 0.00447512 0.0897066 3.18563 0.768845 0.93505 0.2523 0.147453
Img9 0.00456293 0.0895977 3.54939 0.731027 0.924623 0.243885 0.107223
Img10 0.00469539 0.0898217 3.37518 0.763099 0.933868 0.265854 0.143027

Fig. 8. The comparison using parameters mentioned in Table 2

 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 709

inhomogeneity by standard deviation. A higher Kurtosis (Sk)

value means a higher level of saturation and high Kurtosis is a parameter that describes
contrast at the image’s edges. the form of random variables in probability
distributions. Kurtosis is denoted as Kurt(X) for a
random variable X, and it is described as

...(8)
...(11)
Entropy (E) Entropy is a measure of the
Energy (En)
random of a textural picture that is calculated and
The energy is a measurable total of the
expressed as
degree of repetitions for pair pixels. Energy is a
metric for evaluating how close two images. If
Haralick’s GLCM function is used to define energy,
it is also known as pointed moment, and it exists
expressed as
...(9)
Skewness (Sk)
Skewness is a measurement of absence
there. (X) is the random variable X, and it is defined
as ...(12)
Contrast (Con)
It is defined as the pixel’s intensity and
...(10) that of its neighbors across a contrasting picture,
and it is described as
Table 3. TP, TN, FP and FN used for Confusion Matrix
terms

Result Negative Positive Total

Negative TN FN TN+FN
Positive FP TP FP+TP
Total FP+TN TP+FN FP+TN+ TP+FN

Table 4. Training and validation images preparing (5-fold cross-validation)

Total no. Training (128) Validation (32)

of images Normal Abnormal Normal Abnormal

160 16 112 4 28

Table 5. Parameters for segmented tissue performance

Methods MSE PSNR SSIM Dice score

K-NN[25] 0.234 0.75 0.8761 0.81

SOM[24] 0.21 0.76 0.9101 0.83
GA[26] 0.026 0.78 0.9706 0.85
GCNN[28] 0.0324 0.89 0.9863 0.89
Proposed method 0.0012 0.98 0.9901 0.94
710 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

...(15)
...(13) Correlation (Corr )
Inverse Difference Moment (IDM) or The spatial dependencies between pixels
Homogeneity are described by the correlation function, which is
Inverse difference the limited homogeneity defined as
of an image is considered with its moment. To
decide whether an image is textured or not textured,
IDM can have a single value or multiple values.
...(16)
Coarseness (Cness )
In image textural analysis, coarseness
is a roughness measure. The coarser the texture,
...(14) the higher the coarseness score window size.
Directional Moment (DM) Coarseness values are lower in fine textures. It is
It is described as a measurement of the defined as
textural motilities picture utilizing the image’s
alignment as a metric in terms of angle.

...(17)

Table 6. Compassion of accuracies in different clarifiers methodology.

Methods TP FP TN FN Precision Recall Accuracy

K-NN[25] 0.8923 1.1441 0.851 1.3843 0.7974 0.9472 0.8503

SOM[24] 0.9213 0.8122 0.893 1.2123 0.8054 0.9654 0.9154
GA[26] 0.9342 0.7642 0.924 0.8765 0.8443 0.9586 0.9767
GCNN[28] 0.9423 0.1123 0.942 0.1654 0.9324 0.9652 0.9589
Proposed method 0.9854 0.0121 0.985 0.0176 0.9543 0.9765 0.9875

Fig. 9. The represent of graphical comparison through the parameters mention in Table 4
 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 711

Classification The most frequently used kernel functions

Kernel-based Support Vector Machine (KSVM) are off-linear kernels, quadratic kernels, polynomial
It’s a good tool for data processing kernels, and sigmoid kernels. The expressions for
and classification, and it’s based on supervised kernel functions are as follows:
learning. It has a fast-learning speed even when Linear kernel
dealing with large amounts of data. It is used to
solve grouping problems in two or more groups.
It is based largely on the division plane theory. A
division plane is a plane that divides a series of data ...(18)
into groups of separate memberships. It is used in Quadratic kernel
the identification and classification of MR brain
tumors. The SVM algorithm is used to classify the
tumor class shown in the graphic. It consists of two
parts: training and testing. ...(19)

Fig. 10. The represent of graphical comparison through the parameters mention in Table 6

Fig. 11. Accuracy comparison of proposed and existing methodology

712 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

Polynomial kernel was chosen because, in comparison to the T1 and

PET modalities, T2 images had more contrast
...(20) and clarity. For the experiment, 160 patients’
MRIs were chosen, including 20 normal and 140
d >0 is the degree of the polynomial. abnormal MRIs, as there is one kind of normal
Sigmoid kernel brain MRIs and seven types of abnormal brain
MRIs in the dataset. Further, 132 cases have been
used for training and 28 cases for validation in the
...(21) 5-fold cross-validation, the outcomes have been
RBF kernel shown in Table 4.
K-Fold Cross-validation
Subsequently, the trained classifier
...(22) on a single dataset can only succeed with high
Where px, py represents the inner classification precision for that dataset and not
products of the kernel and ‘z’ is a constant. for additional datasets. Cross-validation must be
Experimental Results And Discussion incorporated into our system to avoid overfitting.
This section presents a detailed discussion Cross-validation will not improve the classification
of the result analysis using both synthetic and actual accuracy in the end; however, it will make the
brain MRI images. Experiments were carried out classifier more accurate and generalizable to other
using MATLAB tool, which was installed in a PC datasets.
with Windows 10 OS, 4 GB of RAM, and an Intel Segmentation Performance Measures
Core i5 processor. To compare the effectiveness The suggested unsupervised clustering
of the proposed framework, four state-of-art algorithms for the segmentation of brain tumors
methods are used such as K-NN25, SOM24, GA26, are tested using the following performance tests.
and GCNN28. Mean squared error(MSE)
Dataset The MSE (mean squared error) is a
MRI images of T2-weighted with measure of how well everything does. MSE
256×256 in-plane resolution in the axial plane indicates the mechanisms of squaring the distinct
were obtained from The AANLIB dataset for principles. The MSE is the average of the number
brain images (http://adni.loni.ucla.edu/), the of squares of the defects, which is calculated by
OASIS dataset for brain images (http://www. subtracting the input image segmented (Fig. 6).
oasis-brains.org/), and Harvard Medical School MSE may be written as follows:
(http://med.harvard.edu/AANLIB/). The T2 model

Fig. 12. Computation time in seconds for comparative analysis

 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 713

...(26)
...(23)
Where R (a, b) and S (a, b) are respectively Where A∈ {0, 1} denotes the tumor area
input and segmented image. The total rows and as determined by algorithms and B∈ {0, 1} is the
columns from MRI brain image input are m and ground truth as determined by experts. The dice
n. For a better PSNR, the obtained MSE value for coefficient has a minimum and maximum 0, 1
the segmented image should be tiny. respectively; a higher number indicates that the
PSNR (peak signal to noise ratio) two images overlap more.
The peak signal to noise ratio shows an Performance analysis
image’s noise immunity. More PSNR shows that The features of randomly selected 10
noise disturbance in the MRI brain image is low. benign samples are shown in Table 1. Fig. 7
The largest pixel value in the input MR brain image shows the result for Graphical representation by
is designated as a maxi. the parameters mentioned in Table 1. In Table 2,
The PSNR is defined by the MSE the features of randomly selected 10 malignant are
values. PSNR values in the range of 40 to 100 dB presented. The graphical representation based on
were produced by the algorithm, making it less parameters mentioned in Table 2 is shown in Fig
susceptible to noise. PSNR is commonly expressed 8.
as Table 3 shows the definition of the terms
for the confusion matrix as TP, TN, FP, and FN.
Table 4 shows 5-fold cross-validation
using and their training and validation set to the
images.
...(24) Table 5 shows the value of MSE, PSNR,
and SSIM different techniques.
Structural Similarity Index (SSIM) The graphical findings for the mentioned
Structural Similarity Index (SSIM) stands table 4 are seen in Fig. 9. It shows that, as compared
as a perceptual metric that shows when picture to the approaches described above, the designed
quality is degraded due to data compression, data approach yields successful results.
transmission losses, or another image processing The values of precision and recall various
approach. It’s described as methods are shown in Table 6 as TP, FP, TN, and
FN.
The graphical findings for the mentioned
table 5 are seen in Fig. 10. It was found that the
proposed technique outperformed other approaches.
The suggested approach is compared to K-NN,
...(25) SOM, GN, and CNN methods. Fig. 11 shows the
graphical performance accuracy comparison of
An advanced SSIM value means better proposed and existing methodologies.
luminance, contrast, and structural material Fig. 12 shows the graphical performance
preservation. computation time in seconds for comparative
analysis.
Dice Coefficient TP (True Positive): Present tumor that was
The dice coefficient, also known as the appropriately identified.
dice similarity index, is an indicator of when the TN (True Negative): Tumor that does not
two images compare, and it is expressed as occur and was not detected.
FP (False Positive): recognition of a non-
existing tumor.
714 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

FN (False Negative): Nearby is an existing Conclusion and Future Work

tumor that has not been detected. Precision is a
metric for determining whether a person has a This paper has proposed an efficient
tumor. automated brain tumor segmentation and
Precision is given by: classification framework using K-means clustering
and kernel-based support vector machine
(K-SVM). The major steps of this work consist of
...(27) preprocessing, segmentation, feature extraction
including selection, and classification for brain
The measure of recall is the successful tumors. The regions of interest (ROI) are extracted
determination of the person not having a tumor. using skull stripping and median filter. Discrete
Wavelet Transform (DWT)-based texture features
The specificity is given by: are used to feature extraction and significant
features are selected by the principal component
analysis (PCA). Kernel-based support vector
...(28) machine (K-SVM) is applied for the classification
of brain tumor types into benign and malignant.
The measure of accuracy is the successful Result analysis confirms the efficiency of the
classification. proposed framework for classifying the benign or
malignant tissues using brain MRI images, with
The accuracy is given by: 98.75% accuracy, 95.43% precision, and 97.65%
recall. The simulation findings emphasize the
importance of the proposed approach is compared
to state-of-the-art techniques, in terms of accuracy,
...(29) precision, and recall. In future work, we plan
to evaluate the classifier’s selective scheme by
Discussion integrating several classifiers and feature selection
processes.
The proposed approach is quite likely
to create more effective results than the existing Acknowledgement
mentioned techniques. The following are some of
the outcomes of the experiment: The authors would like to thank Dr.
The suggested method extracts more Yashwant Kurmi, Madhyanchal Professional
information to characterize the ROI more properly. University Bhopal M.P. for the necessary support
The recommended technique produces to do this study.
significantly better results than the other techniques Conflict of Interest
mentioned. The authors complete not have any
The image quality has not been degraded conflict of interest.
by the proposed technique. Funding Source
Future development in the approach could This learning received no external
involve the following: funding.
The proposed work may be tested and
extended for other public databases such as References
PASCAL, Berkeley, or BRATS.
To develop the performance of the 1. Guo, Lei, et al. “Tumor detection in MR images
using one-class immune feature weighted
proposed method, some hybrid approaches with
SVMs.” IEEE Transactions on Magnetics, 2011;
different classifiers can be tested in the future.
47(10): 3849-3852.
 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022) 715

2. Kumari, Rosy. “SVM classification an approach 14. Poteraº CM, Mihãescu MC, Mocanu M. An
on detecting abnormality in brain MRI optimized version of the K-Means clustering
images.” International Journal of Engineering algorithm. In 2014 Federated Conference on
Research and Applications, 2013; 3(4): 1686- Computer Science and Information Systems,
1690. 2014(pp. 695-699). IEEE.
3. Gordillo, Nelly, Eduard Montseny, and Pilar 15. Alfonse, Marco, and Abdel-Badeeh M. Salem.
Sobrevilla. “State of the art survey on MRI “An automatic classification of brain tumors
brain tumor segmentation.” Magnetic resonance through MRI using support vector machine.” Egy.
imaging, 2013; 31(8): 1426-1438. Comp. Sci. J, 2016; 40(3).
4. Demirhan, Ayþe, Mustafa Törü, and ÝnanGüler. 16. Yogalakshmi, G., and B. Sheela Rani.
“Segmentation of tumor and edema along with “A Review On The Techniques Of Brain
healthy tissues of brain using wavelets and neural Tumor: Segmentation, Feature Extraction
networks.” IEEE journal of biomedical and And Classification.” 2020 11th International
health informatics, 2014; 19(4): 1451-1458. Conference on Computing, Communication
5. Hamad, Yousif Ahmed, Konstantin Vasilievich and Networking Technologies (ICCCNT). IEEE,
S i m o n o v, a n d M o h a m m a d B . N a e e m . 2020.
“Detection of brain tumor in MRI images, 17. Sharif, Muhammad, et al. “An integrated design
using a combination of fuzzy C-means and of particle swarm optimization (PSO) with fusion
thresholding.” International Journal of Advanced of features for detection of brain tumor.” Pattern
Pervasive and Ubiquitous Computing (IJAPUC), Recognition Letters, 2020; 129: 150-157.
2019; 11(1): 45-60. 18. Moeskops, Pim, et al. “Adversarial training
6. Singh, Garima, and M. A. Ansari. “Efficient and dilated convolutions for brain MRI
detection of brain tumor from MRIs using segmentation.” Deep learning in medical image
K-means segmentation and normalized analysis and multimodal learning for clinical
histogram.” 2016 1st India International decision support. Springer, Cham, 2017; 56-64.
Conference on Information Processing (IICIP). 19. Sahu, Satya Prakash, et al. “Segmentation
IEEE, (2016). of Lungs in Thoracic CTs Using K-means
7. Shree, N. Varuna, and T. N. R. Kumar. Clustering and Morphological Operations.”
“Identification and classification of brain tumor Advances in Biomedical Engineering and
MRI images with feature extraction using Technology. Springer, Singapore, 2021; 331-343.
DWT and probabilistic neural network.” Brain 20. Menze, Bjoern H., et al. “The multimodal
informatics, 2018; 5(1): 23-30. brain tumor image segmentation benchmark
8. Pereira, Sérgio, et al. “Brain tumor segmentation (BRATS).” IEEE transactions on medical
using convolutional neural networks in MRI imaging, 2014; 34(10): 1993-2024.
images.” IEEE transactions on medical imaging, 21. Prastawa, Marcel, et al. “A brain tumor
2016; 35(5): 1240-1251. segmentation framework based on outlier
9. Deepa, S. N., And B. Arunadevi. “Extreme detection.” Medical image analysis, 2004; 8(3):
Learning Machine For Classification 275-283.
Of Brain Tumor In 3d Mr Images/Elm Za 22. Singh, Amritpal. “Detection of brain tumor
Klasifikaciju Tumora Mozga Kod 3d Mr in MRI images, using combination of fuzzy
Snimaka.” Informatologia, 2013; 46(2): 111. c-means and SVM.” 2015 2nd International
10. Sachdeva, Jainy, et al. “Segmentation, Conference on Signal Processing and Integrated
feature extraction, and multiclass brain tumor Networks (SPIN). IEEE, 2015.
classification.” Journal of digital imaging, 2013; 23. Hiremath, P. S., S. Shivashankar, and Jagadeesh
26(6): 1141-1150. Pujari. “Wavelet based features for color
11. Zotin, Alexander, et al. “Edge detection in MRI texture classification with application to
brain tumor images based on fuzzy C-means CBIR.” International Journal of Computer
clustering.” Procedia Computer Science, 2018; Science and Network Security, 2006; 6(9A):
126: 1261-1270. 124-133.
12. Li, Ming, et al. “Brain tumor detection based on 24. Logeswari, T., and M. Karnan. “An improved
multimodal information fusion and convolutional implementation of brain tumor detection
neural network.” IEEE Access, 2019; 7: 180134- using segmentation based on hierarchical self
180146. organizing map.” International Journal of
13. Huang, Hong, et al. “Brain image segmentation Computer Theory and Engineering, 2010; 2(4):
based on FCM clustering algorithm and rough 591.
set.” IEEE Access, 2019; 7: 12386-12396. 25. Vrooman, Henri A., et al. “Multi-spectral brain
716 Mandle et al., Biomed. & Pharmacol. J, Vol. 15(2), 699-716 (2022)

tissue segmentation using automatically trained 27. Ouseph, N. C., and K. Shruti. “A reliable
k-Nearest-Neighbor classification.” Neuroimage, method for brain tumor detection using cnn
2007; 37(1): 71-81. technique.” National Conference on Emerging
26. Kharrat, Ahmed, et al. “A hybrid approach Research Trends in Electrical, Electronics &
for automatic classification of brain MRI Instrumentation (ERTEEI’17). 2017.
using genetic algorithm and support vector 28. Mittal, Mamta, et al. “Deep learning based
machine.” Leonardo journal of sciences, 2010; enhanced tumor segmentation approach for
17(1): 71-82. MR brain images.” Applied Soft Computing 78
(2019): 346-354.